Loading ToolBench
MCP server providing AI tools for code execution, memory management, web search, and deep research integration
IronHive Host provides 12 tools with mixed quality. Naming is generally verb-forward (code_execute, memory_store, web_search), but several tools lack complete input schemas visible in source. Descriptions are present but often generic (e.g., memory_recall: 'Recalls memories semantically related to the query', 72 chars, below the 194-char baseline). Parameter descriptions exist but are minimal. Output schemas are not documented in the source code provided. Error handling is basic (try-catch with generic error messages). Tools like mcp_discover_tools and mcp_load_plugin are inferred from test files only, not explicitly registered in production code, capping their scores at 50. Code execution tools (code_execute, code_session, code_install_packages) have reasonable structure but lack output schema documentation and error recovery guidance.
Executes code in a sandboxed environment. Supports: javascript, typescript, python, csharp. Returns stdout, stderr, and execution result.
Installs packages in an execution session. Provide package names as an array (e.g., ['lodash', 'axios'] for JavaScript).
Manages code execution sessions. Actions: 'create' (new session), 'destroy' (end session), 'list' (show active sessions). Sessions maintain state between executions.
Perform deep autonomous research on a topic. Iteratively searches the web, analyzes sources, and generates a comprehensive research report with citations. Use this for complex research queries that require multiple rounds of investigation.
Explore a website's structure by analyzing its robots.txt and sitemap. Returns URL list and site structure. Use this to systematically discover content on a specific site.
Discovers available MCP tools from configured plugins
Output schemas not documented. Tools return JSON-serialized results (e.g., ExecutionResult, SessionResult) but the schema structure is not visible in tool definitions. LLMs cannot plan downstream calls or extract fields reliably.
mcp_discover_tools and mcp_load_plugin are inferred from test files only (tests/IronHive.Host.Tests/Agent/Mcp/McpToolDiscoveryTests.cs), not explicitly registered in production code. No input schemas visible.
Error handling is generic. Code execution tools catch exceptions and return {Success: false, Error: ex.Message}, but do not categorize errors as retryable, user-fixable, or fatal. No recovery guidance provided to LLM.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 54 | 2026-07-28+ | v2 |
Loads an MCP plugin and makes its tools available
Removes a specific memory by its ID. Use with caution as this cannot be undone.
Recalls memories semantically related to the query. Returns the most relevant memories based on vector similarity.
Searches memories with optional filters. Supports filtering by category, tier (Buffer/Short/Long/Archive), and semantic query.
Stores information in long-term memory. Use importance (0.0-1.0) to indicate how critical the information is. Higher importance increases retention.
Search the web for information and return relevant URLs with titles and descriptions. Use this to find up-to-date information or research specific topics.
Parameter descriptions are minimal. E.g., memory_recall 'limit' param: 'Maximum number of results (default: 5)' lacks guidance on when to use different limits or performance implications. Descriptions average ~50 chars; baseline is 72 chars.
No pagination or result-limiting guidance for web_search and deep_research. web_search defaults to 10 results but does not document max cap or whether pagination is supported. Large result sets risk context window exhaustion.